Re: [RFC] printk/sysrq: Don't play with console_loglevel

[Tetsuo Handa]

On 2019/05/29 0:03, Petr Mladek wrote:
>> But is context dependent buffer large enough to hold SysRq-t output?
>> I think that only main logbuf can become large enough to hold SysRq-t output.
> 
> SysRq messages are stored directly into the main log buffer.
> 
> The limited per-CPU buffers are needed only in printk_safe
> and NMI context. We discussed it here because KERN_UNSUPPRESSED
> allows to pass the information even from this context.
> 
>> We can add KERN_UNSUPPRESSED to SysRq's header line. But I don't think
>> that we can automatically add KERN_UNSUPPRESSED to SysRq's body lines
>> based on some context information. If we want to avoid manipulating
>> console_loglevel, we need to think about how to make sure that
>> KERN_UNSUPPRESSED is added to all lines from such context without
>> overflowing capacity of that buffer.
> 
> We could set this context in printk_context per-CPU variable.
> 
> Then we could easily add the set per-message flag in
> vprintk_store() for the normal/atomic context. And we
> could store an extra KERN_UNSUPPRESSED in printk_safe_log_store()
> for printk_safe and NMI context.

Now I got what you are trying to do. Borrow per-CPU printk_context
flags for automatically prefixing KERN_UNSUPPRESSED, based on an
assumption that any message sent during that per-CPU printk_context
flag set is important enough. Then, what we need to be careful is
nesting of setting/clearing of that flag, for NMI handler might be
called during SysRq operation is in progress. We unconditionally
prefix KERN_UNSUPPRESSED if NMI?